Build the intelligence layer from scratch.
A hands-on path covering fine-tuning, RAG, guardrails, agents and feedback loops.
MEDVION
Proof of concept
A 2–3 week path from model to feedback loop.
Based on the supplied OMS AI POC Guide for NVIDIA DGX Spark 128GB, with basic Python listed as the required skill level.
Environment
Python, ML libraries, PostgreSQL/pgvector and Llama 3 8B.
5,000 Examples
Training data across six OMS categories.
Fine-Tune
LoRA adaptation of Llama 3 8B.
RAG
Historical order and resolution retrieval.
Guardrails
Five deterministic safety layers.
Agent
Process mock orders through the stack.
Feedback
Log overrides and retrain with corrections.
Experiments
Measure the system, not just the demo.
The guide proposes comparisons for fine-tuning, RAG, guardrails and feedback-driven retraining.
Base vs fine-tuned
Compare both models on 50 test scenarios.
Context impact
Compare decisions with and without retrieved context.
Safety testing
Feed intentionally bad decisions through guardrails.